Phoneme-to-Audio Forced Alignment with Basic Syllable Types and Broad Phonetic Classes

In this work, an unsupervised system of forced alignment is proposed and can be generalized to any language without the need for manually-aligned training data. Specifically, the phoneme sequences of text transcriptions are transformed into more general phonetic knowledge to guide alignment, namely syllable types and phonetic classes, by using linguistics rules. Instead of using special phoneme recognizer of specific language, the alignment system firstly is solved on the larger and coarser syllable level by using a universal HMM-based syllable model, then the smaller and finer phoneme-level alignment furtherly is implemented within syllable segments by using RNN with CTC. The evaluations are carried out on six corpora of different languages and the results demonstrate the effectiveness of proposed system. Compared to existing state-of-the-art supervised aligners, the proposed system provides competitive performance in the task of forced alignment.